Air Pollution in Five Major Chinese Cities

A widely-circulated 2015 study from American think tank Berkeley Earth estimated from available literature as many as 1.2 to 2 million premature deaths a year from air pollution in China alone. Particulate matter causes “premature death in people with heart and lung disease” per the EPA. These numbers are backed up by the World Health Organization, the Lancet, and the IHME.

Former Health minister Chen Zhu published in the Lancet a conservative estimate of between 350,000 and 500,000 deaths a year. For comparison, Dane, Iowa, Green, and Columbia counties (the Madison Metro statistical area) had a total 2010 population of 568,593.

Reliable, accessible data taken hourly.

Combined population of 5 cities' administrative areas is greater than 100 million people (some 1.5% of the human population and 7% of the population of China)

Air pollution is caused by human activity such as industry, heating, and transportation and cities are one obvious starting point for China's efforts to improve air quality.

The United States Department of State measures air quality data at five locations–the U.S. Embassy in Beijing, and the Consulates in Chengdu, Guangzhou, Shanghai, and Shenyang.

The data is measured in terms of PM2.5, which measures the total mass of fine particles in the atmosphere (micrograms per cubic meter).

While PM2.5 is caused by human activity, we think an intriguing area of study is the short-term effect of meterological data for the five cities on the readings.

This data (from the site “Weather Underground”) contains temperature, dew-point, humidity, sea-level pressure, visibility, and wind speed.

• Step 1 Combine meteorological data with the Mean.PM2.5 data.
• Step 2 Fit regression model without days indicate significant high values of PM2.5(>250).
• Step 3 Do specific reaserch&analysis for days that PM2.5>250 (outliers).

Parade Blue demonstrated how sensitive particulate readings are to short-term human activities (and reduction).

We're using hours 0, 8, 12, 18 in order to account for the effect of the most important times of day without overcomplicating things.

How do we deal with missing values?

(delete 1483 missing values [after sending them to -999])

• Step 1 Combine meteorological data with the Mean.PM2.5 data.
• Step 2 Fit regression model without days indicate significant high values of PM2.5(>250).
• Step 3 Do specific reaserch&analysis for days that PM2.5>250 (outliers).

Next step:Check if there is a similar pattern across the years Find factors related to outliers

Some blind alleys

NOAA Reanalysis dataset–perfect, but way too complicated to interpret, parse, and group. Simpler datasets are available.

Lots of missing data–meant for building climate models, not small-scale, localized projects.

“Self-documenting code”, but requires some expert domain knowledge

Idea is to treat city as a fixed effect and compute readings by month/day as random effects with multiple error terms

Makes more sense for grouping dozens of messy, missing, spatially-dispersed meteorogical data with multiple readings–not so useful for estimating parameters with single samples.

Might be useful for handling “repeated samples” from groupings in mornings, afternoon, evening, night.

More useful for ggplot/organization than for stats in this project (1 object instead of 5)